If we are doing direct IO writes with datasync semantics, we often
have to flush metadata changes along with the data write. However,
if we are overwriting existing data, there are no metadata changes
that we need to flush. In this case, optimising the IO by using
FUA write makes sense.

We know from the IOMAP_F_DIRTY flag as to whether a specific inode
requires a metadata flush - this is currently used by DAX to ensure
extent modification as stable in page fault operations. For direct
IO writes, we can use it to determine if we need to flush metadata
or not once the data is on disk.

Hence if we have been returned a mapped extent that is not new and
the IO mapping is not dirty, then we can use a FUA write to provide
datasync semantics. This allows us to short-cut the
generic_write_sync() call in IO completion and hence avoid
unnecessary operations. This makes pure direct IO data write
behaviour identical to the way block devices use REQ_FUA to provide
datasync semantics.

On a FUA enabled device, a synchronous direct IO write workload
(sequential 4k overwrites in 32MB file) had the following results:

# xfs_io -fd -c "pwrite -V 1 -D 0 32m" /mnt/scratch/boo

kernel		time	write()s	write iops	Write b/w
------		----	--------	----------	---------
(no dsync)	 4s	2173/s		2173		8.5MB/s
vanilla		22s	 370/s		 750		1.4MB/s
patched		19s	 420/s		 420		1.6MB/s

The patched code clearly doesn't send cache flushes anymore, but
instead uses FUA (confirmed via blktrace), and performance improves
a bit as a result. However, the benefits will be higher on workloads
that mix O_DSYNC overwrites with other write IO as we won't be
flushing the entire device cache on every DSYNC overwrite IO
anymore.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>